1 /*
2  * Copyright 2011 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <deathsimple@vodafone.de>
29  */
30 
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 
34 #include <drm/drm.h>
35 #include <drm/drm_drv.h>
36 
37 #include "amdgpu.h"
38 #include "amdgpu_pm.h"
39 #include "amdgpu_uvd.h"
40 #include "cikd.h"
41 #include "uvd/uvd_4_2_d.h"
42 
43 #include "amdgpu_ras.h"
44 
45 /* 1 second timeout */
46 #define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
47 
48 /* Firmware versions for VI */
49 #define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
50 #define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
51 #define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
52 #define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
53 
54 /* Polaris10/11 firmware version */
55 #define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
56 
57 /* Firmware Names */
58 #ifdef CONFIG_DRM_AMDGPU_SI
59 #define FIRMWARE_TAHITI		"amdgpu/tahiti_uvd.bin"
60 #define FIRMWARE_VERDE		"amdgpu/verde_uvd.bin"
61 #define FIRMWARE_PITCAIRN	"amdgpu/pitcairn_uvd.bin"
62 #define FIRMWARE_OLAND		"amdgpu/oland_uvd.bin"
63 #endif
64 #ifdef CONFIG_DRM_AMDGPU_CIK
65 #define FIRMWARE_BONAIRE	"amdgpu/bonaire_uvd.bin"
66 #define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
67 #define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
68 #define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
69 #define FIRMWARE_MULLINS	"amdgpu/mullins_uvd.bin"
70 #endif
71 #define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
72 #define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
73 #define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
74 #define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
75 #define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
76 #define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
77 #define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
78 #define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
79 
80 #define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
81 #define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
82 #define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
83 
84 /* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
85 #define UVD_GPCOM_VCPU_CMD		0x03c3
86 #define UVD_GPCOM_VCPU_DATA0	0x03c4
87 #define UVD_GPCOM_VCPU_DATA1	0x03c5
88 #define UVD_NO_OP				0x03ff
89 #define UVD_BASE_SI				0x3800
90 
91 /*
92  * amdgpu_uvd_cs_ctx - Command submission parser context
93  *
94  * Used for emulating virtual memory support on UVD 4.2.
95  */
96 struct amdgpu_uvd_cs_ctx {
97 	struct amdgpu_cs_parser *parser;
98 	unsigned reg, count;
99 	unsigned data0, data1;
100 	unsigned idx;
101 	unsigned ib_idx;
102 
103 	/* does the IB has a msg command */
104 	bool has_msg_cmd;
105 
106 	/* minimum buffer sizes */
107 	unsigned *buf_sizes;
108 };
109 
110 #ifdef CONFIG_DRM_AMDGPU_SI
111 MODULE_FIRMWARE(FIRMWARE_TAHITI);
112 MODULE_FIRMWARE(FIRMWARE_VERDE);
113 MODULE_FIRMWARE(FIRMWARE_PITCAIRN);
114 MODULE_FIRMWARE(FIRMWARE_OLAND);
115 #endif
116 #ifdef CONFIG_DRM_AMDGPU_CIK
117 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
118 MODULE_FIRMWARE(FIRMWARE_KABINI);
119 MODULE_FIRMWARE(FIRMWARE_KAVERI);
120 MODULE_FIRMWARE(FIRMWARE_HAWAII);
121 MODULE_FIRMWARE(FIRMWARE_MULLINS);
122 #endif
123 MODULE_FIRMWARE(FIRMWARE_TONGA);
124 MODULE_FIRMWARE(FIRMWARE_CARRIZO);
125 MODULE_FIRMWARE(FIRMWARE_FIJI);
126 MODULE_FIRMWARE(FIRMWARE_STONEY);
127 MODULE_FIRMWARE(FIRMWARE_POLARIS10);
128 MODULE_FIRMWARE(FIRMWARE_POLARIS11);
129 MODULE_FIRMWARE(FIRMWARE_POLARIS12);
130 MODULE_FIRMWARE(FIRMWARE_VEGAM);
131 
132 MODULE_FIRMWARE(FIRMWARE_VEGA10);
133 MODULE_FIRMWARE(FIRMWARE_VEGA12);
134 MODULE_FIRMWARE(FIRMWARE_VEGA20);
135 
136 static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
137 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo);
138 
139 static int amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device *adev,
140 					   uint32_t size,
141 					   struct amdgpu_bo **bo_ptr)
142 {
143 	struct ttm_operation_ctx ctx = { true, false };
144 	struct amdgpu_bo *bo = NULL;
145 	void *addr;
146 	int r;
147 
148 	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
149 				      AMDGPU_GEM_DOMAIN_GTT,
150 				      &bo, NULL, &addr);
151 	if (r)
152 		return r;
153 
154 	if (adev->uvd.address_64_bit)
155 		goto succ;
156 
157 	amdgpu_bo_kunmap(bo);
158 	amdgpu_bo_unpin(bo);
159 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
160 	amdgpu_uvd_force_into_uvd_segment(bo);
161 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
162 	if (r)
163 		goto err;
164 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
165 	if (r)
166 		goto err_pin;
167 	r = amdgpu_bo_kmap(bo, &addr);
168 	if (r)
169 		goto err_kmap;
170 succ:
171 	amdgpu_bo_unreserve(bo);
172 	*bo_ptr = bo;
173 	return 0;
174 err_kmap:
175 	amdgpu_bo_unpin(bo);
176 err_pin:
177 err:
178 	amdgpu_bo_unreserve(bo);
179 	amdgpu_bo_unref(&bo);
180 	return r;
181 }
182 
183 int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
184 {
185 	unsigned long bo_size;
186 	const char *fw_name;
187 	const struct common_firmware_header *hdr;
188 	unsigned family_id;
189 	int i, j, r;
190 
191 	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
192 
193 	switch (adev->asic_type) {
194 #ifdef CONFIG_DRM_AMDGPU_SI
195 	case CHIP_TAHITI:
196 		fw_name = FIRMWARE_TAHITI;
197 		break;
198 	case CHIP_VERDE:
199 		fw_name = FIRMWARE_VERDE;
200 		break;
201 	case CHIP_PITCAIRN:
202 		fw_name = FIRMWARE_PITCAIRN;
203 		break;
204 	case CHIP_OLAND:
205 		fw_name = FIRMWARE_OLAND;
206 		break;
207 #endif
208 #ifdef CONFIG_DRM_AMDGPU_CIK
209 	case CHIP_BONAIRE:
210 		fw_name = FIRMWARE_BONAIRE;
211 		break;
212 	case CHIP_KABINI:
213 		fw_name = FIRMWARE_KABINI;
214 		break;
215 	case CHIP_KAVERI:
216 		fw_name = FIRMWARE_KAVERI;
217 		break;
218 	case CHIP_HAWAII:
219 		fw_name = FIRMWARE_HAWAII;
220 		break;
221 	case CHIP_MULLINS:
222 		fw_name = FIRMWARE_MULLINS;
223 		break;
224 #endif
225 	case CHIP_TONGA:
226 		fw_name = FIRMWARE_TONGA;
227 		break;
228 	case CHIP_FIJI:
229 		fw_name = FIRMWARE_FIJI;
230 		break;
231 	case CHIP_CARRIZO:
232 		fw_name = FIRMWARE_CARRIZO;
233 		break;
234 	case CHIP_STONEY:
235 		fw_name = FIRMWARE_STONEY;
236 		break;
237 	case CHIP_POLARIS10:
238 		fw_name = FIRMWARE_POLARIS10;
239 		break;
240 	case CHIP_POLARIS11:
241 		fw_name = FIRMWARE_POLARIS11;
242 		break;
243 	case CHIP_POLARIS12:
244 		fw_name = FIRMWARE_POLARIS12;
245 		break;
246 	case CHIP_VEGA10:
247 		fw_name = FIRMWARE_VEGA10;
248 		break;
249 	case CHIP_VEGA12:
250 		fw_name = FIRMWARE_VEGA12;
251 		break;
252 	case CHIP_VEGAM:
253 		fw_name = FIRMWARE_VEGAM;
254 		break;
255 	case CHIP_VEGA20:
256 		fw_name = FIRMWARE_VEGA20;
257 		break;
258 	default:
259 		return -EINVAL;
260 	}
261 
262 	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
263 	if (r) {
264 		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
265 			fw_name);
266 		return r;
267 	}
268 
269 	r = amdgpu_ucode_validate(adev->uvd.fw);
270 	if (r) {
271 		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
272 			fw_name);
273 		release_firmware(adev->uvd.fw);
274 		adev->uvd.fw = NULL;
275 		return r;
276 	}
277 
278 	/* Set the default UVD handles that the firmware can handle */
279 	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
280 
281 	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
282 	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
283 
284 	if (adev->asic_type < CHIP_VEGA20) {
285 		unsigned version_major, version_minor;
286 
287 		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
288 		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
289 		DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
290 			version_major, version_minor, family_id);
291 
292 		/*
293 		 * Limit the number of UVD handles depending on microcode major
294 		 * and minor versions. The firmware version which has 40 UVD
295 		 * instances support is 1.80. So all subsequent versions should
296 		 * also have the same support.
297 		 */
298 		if ((version_major > 0x01) ||
299 		    ((version_major == 0x01) && (version_minor >= 0x50)))
300 			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
301 
302 		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
303 					(family_id << 8));
304 
305 		if ((adev->asic_type == CHIP_POLARIS10 ||
306 		     adev->asic_type == CHIP_POLARIS11) &&
307 		    (adev->uvd.fw_version < FW_1_66_16))
308 			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
309 				  version_major, version_minor);
310 	} else {
311 		unsigned int enc_major, enc_minor, dec_minor;
312 
313 		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
314 		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
315 		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
316 		DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
317 			enc_major, enc_minor, dec_minor, family_id);
318 
319 		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
320 
321 		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
322 	}
323 
324 	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
325 		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
326 	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
327 		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
328 
329 	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
330 		if (adev->uvd.harvest_config & (1 << j))
331 			continue;
332 		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
333 					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
334 					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
335 		if (r) {
336 			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
337 			return r;
338 		}
339 	}
340 
341 	for (i = 0; i < adev->uvd.max_handles; ++i) {
342 		atomic_set(&adev->uvd.handles[i], 0);
343 		adev->uvd.filp[i] = NULL;
344 	}
345 
346 	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
347 	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
348 		adev->uvd.address_64_bit = true;
349 
350 	r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
351 	if (r)
352 		return r;
353 
354 	switch (adev->asic_type) {
355 	case CHIP_TONGA:
356 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
357 		break;
358 	case CHIP_CARRIZO:
359 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
360 		break;
361 	case CHIP_FIJI:
362 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
363 		break;
364 	case CHIP_STONEY:
365 		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
366 		break;
367 	default:
368 		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
369 	}
370 
371 	return 0;
372 }
373 
374 int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
375 {
376 	void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
377 	int i, j;
378 
379 	drm_sched_entity_destroy(&adev->uvd.entity);
380 
381 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
382 		if (adev->uvd.harvest_config & (1 << j))
383 			continue;
384 		kvfree(adev->uvd.inst[j].saved_bo);
385 
386 		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
387 				      &adev->uvd.inst[j].gpu_addr,
388 				      (void **)&adev->uvd.inst[j].cpu_addr);
389 
390 		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
391 
392 		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
393 			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
394 	}
395 	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
396 	release_firmware(adev->uvd.fw);
397 
398 	return 0;
399 }
400 
401 /**
402  * amdgpu_uvd_entity_init - init entity
403  *
404  * @adev: amdgpu_device pointer
405  *
406  */
407 int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
408 {
409 	struct amdgpu_ring *ring;
410 	struct drm_gpu_scheduler *sched;
411 	int r;
412 
413 	ring = &adev->uvd.inst[0].ring;
414 	sched = &ring->sched;
415 	r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
416 				  &sched, 1, NULL);
417 	if (r) {
418 		DRM_ERROR("Failed setting up UVD kernel entity.\n");
419 		return r;
420 	}
421 
422 	return 0;
423 }
424 
425 int amdgpu_uvd_suspend(struct amdgpu_device *adev)
426 {
427 	unsigned size;
428 	void *ptr;
429 	int i, j, idx;
430 	bool in_ras_intr = amdgpu_ras_intr_triggered();
431 
432 	cancel_delayed_work_sync(&adev->uvd.idle_work);
433 
434 	/* only valid for physical mode */
435 	if (adev->asic_type < CHIP_POLARIS10) {
436 		for (i = 0; i < adev->uvd.max_handles; ++i)
437 			if (atomic_read(&adev->uvd.handles[i]))
438 				break;
439 
440 		if (i == adev->uvd.max_handles)
441 			return 0;
442 	}
443 
444 	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
445 		if (adev->uvd.harvest_config & (1 << j))
446 			continue;
447 		if (adev->uvd.inst[j].vcpu_bo == NULL)
448 			continue;
449 
450 		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
451 		ptr = adev->uvd.inst[j].cpu_addr;
452 
453 		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
454 		if (!adev->uvd.inst[j].saved_bo)
455 			return -ENOMEM;
456 
457 		if (drm_dev_enter(adev_to_drm(adev), &idx)) {
458 			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
459 			if (in_ras_intr)
460 				memset(adev->uvd.inst[j].saved_bo, 0, size);
461 			else
462 				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
463 
464 			drm_dev_exit(idx);
465 		}
466 	}
467 
468 	if (in_ras_intr)
469 		DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
470 
471 	return 0;
472 }
473 
474 int amdgpu_uvd_resume(struct amdgpu_device *adev)
475 {
476 	unsigned size;
477 	void *ptr;
478 	int i, idx;
479 
480 	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
481 		if (adev->uvd.harvest_config & (1 << i))
482 			continue;
483 		if (adev->uvd.inst[i].vcpu_bo == NULL)
484 			return -EINVAL;
485 
486 		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
487 		ptr = adev->uvd.inst[i].cpu_addr;
488 
489 		if (adev->uvd.inst[i].saved_bo != NULL) {
490 			if (drm_dev_enter(adev_to_drm(adev), &idx)) {
491 				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
492 				drm_dev_exit(idx);
493 			}
494 			kvfree(adev->uvd.inst[i].saved_bo);
495 			adev->uvd.inst[i].saved_bo = NULL;
496 		} else {
497 			const struct common_firmware_header *hdr;
498 			unsigned offset;
499 
500 			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
501 			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
502 				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
503 				if (drm_dev_enter(adev_to_drm(adev), &idx)) {
504 					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
505 						    le32_to_cpu(hdr->ucode_size_bytes));
506 					drm_dev_exit(idx);
507 				}
508 				size -= le32_to_cpu(hdr->ucode_size_bytes);
509 				ptr += le32_to_cpu(hdr->ucode_size_bytes);
510 			}
511 			memset_io(ptr, 0, size);
512 			/* to restore uvd fence seq */
513 			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
514 		}
515 	}
516 	return 0;
517 }
518 
519 void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
520 {
521 	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
522 	int i, r;
523 
524 	for (i = 0; i < adev->uvd.max_handles; ++i) {
525 		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
526 
527 		if (handle != 0 && adev->uvd.filp[i] == filp) {
528 			struct dma_fence *fence;
529 
530 			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
531 						       &fence);
532 			if (r) {
533 				DRM_ERROR("Error destroying UVD %d!\n", r);
534 				continue;
535 			}
536 
537 			dma_fence_wait(fence, false);
538 			dma_fence_put(fence);
539 
540 			adev->uvd.filp[i] = NULL;
541 			atomic_set(&adev->uvd.handles[i], 0);
542 		}
543 	}
544 }
545 
546 static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
547 {
548 	int i;
549 	for (i = 0; i < abo->placement.num_placement; ++i) {
550 		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
551 		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
552 	}
553 }
554 
555 static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
556 {
557 	uint32_t lo, hi;
558 	uint64_t addr;
559 
560 	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
561 	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
562 	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
563 
564 	return addr;
565 }
566 
567 /**
568  * amdgpu_uvd_cs_pass1 - first parsing round
569  *
570  * @ctx: UVD parser context
571  *
572  * Make sure UVD message and feedback buffers are in VRAM and
573  * nobody is violating an 256MB boundary.
574  */
575 static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
576 {
577 	struct ttm_operation_ctx tctx = { false, false };
578 	struct amdgpu_bo_va_mapping *mapping;
579 	struct amdgpu_bo *bo;
580 	uint32_t cmd;
581 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
582 	int r = 0;
583 
584 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
585 	if (r) {
586 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
587 		return r;
588 	}
589 
590 	if (!ctx->parser->adev->uvd.address_64_bit) {
591 		/* check if it's a message or feedback command */
592 		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
593 		if (cmd == 0x0 || cmd == 0x3) {
594 			/* yes, force it into VRAM */
595 			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
596 			amdgpu_bo_placement_from_domain(bo, domain);
597 		}
598 		amdgpu_uvd_force_into_uvd_segment(bo);
599 
600 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
601 	}
602 
603 	return r;
604 }
605 
606 /**
607  * amdgpu_uvd_cs_msg_decode - handle UVD decode message
608  *
609  * @adev: amdgpu_device pointer
610  * @msg: pointer to message structure
611  * @buf_sizes: placeholder to put the different buffer lengths
612  *
613  * Peek into the decode message and calculate the necessary buffer sizes.
614  */
615 static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
616 	unsigned buf_sizes[])
617 {
618 	unsigned stream_type = msg[4];
619 	unsigned width = msg[6];
620 	unsigned height = msg[7];
621 	unsigned dpb_size = msg[9];
622 	unsigned pitch = msg[28];
623 	unsigned level = msg[57];
624 
625 	unsigned width_in_mb = width / 16;
626 	unsigned height_in_mb = ALIGN(height / 16, 2);
627 	unsigned fs_in_mb = width_in_mb * height_in_mb;
628 
629 	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
630 	unsigned min_ctx_size = ~0;
631 
632 	image_size = width * height;
633 	image_size += image_size / 2;
634 	image_size = ALIGN(image_size, 1024);
635 
636 	switch (stream_type) {
637 	case 0: /* H264 */
638 		switch(level) {
639 		case 30:
640 			num_dpb_buffer = 8100 / fs_in_mb;
641 			break;
642 		case 31:
643 			num_dpb_buffer = 18000 / fs_in_mb;
644 			break;
645 		case 32:
646 			num_dpb_buffer = 20480 / fs_in_mb;
647 			break;
648 		case 41:
649 			num_dpb_buffer = 32768 / fs_in_mb;
650 			break;
651 		case 42:
652 			num_dpb_buffer = 34816 / fs_in_mb;
653 			break;
654 		case 50:
655 			num_dpb_buffer = 110400 / fs_in_mb;
656 			break;
657 		case 51:
658 			num_dpb_buffer = 184320 / fs_in_mb;
659 			break;
660 		default:
661 			num_dpb_buffer = 184320 / fs_in_mb;
662 			break;
663 		}
664 		num_dpb_buffer++;
665 		if (num_dpb_buffer > 17)
666 			num_dpb_buffer = 17;
667 
668 		/* reference picture buffer */
669 		min_dpb_size = image_size * num_dpb_buffer;
670 
671 		/* macroblock context buffer */
672 		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
673 
674 		/* IT surface buffer */
675 		min_dpb_size += width_in_mb * height_in_mb * 32;
676 		break;
677 
678 	case 1: /* VC1 */
679 
680 		/* reference picture buffer */
681 		min_dpb_size = image_size * 3;
682 
683 		/* CONTEXT_BUFFER */
684 		min_dpb_size += width_in_mb * height_in_mb * 128;
685 
686 		/* IT surface buffer */
687 		min_dpb_size += width_in_mb * 64;
688 
689 		/* DB surface buffer */
690 		min_dpb_size += width_in_mb * 128;
691 
692 		/* BP */
693 		tmp = max(width_in_mb, height_in_mb);
694 		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
695 		break;
696 
697 	case 3: /* MPEG2 */
698 
699 		/* reference picture buffer */
700 		min_dpb_size = image_size * 3;
701 		break;
702 
703 	case 4: /* MPEG4 */
704 
705 		/* reference picture buffer */
706 		min_dpb_size = image_size * 3;
707 
708 		/* CM */
709 		min_dpb_size += width_in_mb * height_in_mb * 64;
710 
711 		/* IT surface buffer */
712 		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
713 		break;
714 
715 	case 7: /* H264 Perf */
716 		switch(level) {
717 		case 30:
718 			num_dpb_buffer = 8100 / fs_in_mb;
719 			break;
720 		case 31:
721 			num_dpb_buffer = 18000 / fs_in_mb;
722 			break;
723 		case 32:
724 			num_dpb_buffer = 20480 / fs_in_mb;
725 			break;
726 		case 41:
727 			num_dpb_buffer = 32768 / fs_in_mb;
728 			break;
729 		case 42:
730 			num_dpb_buffer = 34816 / fs_in_mb;
731 			break;
732 		case 50:
733 			num_dpb_buffer = 110400 / fs_in_mb;
734 			break;
735 		case 51:
736 			num_dpb_buffer = 184320 / fs_in_mb;
737 			break;
738 		default:
739 			num_dpb_buffer = 184320 / fs_in_mb;
740 			break;
741 		}
742 		num_dpb_buffer++;
743 		if (num_dpb_buffer > 17)
744 			num_dpb_buffer = 17;
745 
746 		/* reference picture buffer */
747 		min_dpb_size = image_size * num_dpb_buffer;
748 
749 		if (!adev->uvd.use_ctx_buf){
750 			/* macroblock context buffer */
751 			min_dpb_size +=
752 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
753 
754 			/* IT surface buffer */
755 			min_dpb_size += width_in_mb * height_in_mb * 32;
756 		} else {
757 			/* macroblock context buffer */
758 			min_ctx_size =
759 				width_in_mb * height_in_mb * num_dpb_buffer * 192;
760 		}
761 		break;
762 
763 	case 8: /* MJPEG */
764 		min_dpb_size = 0;
765 		break;
766 
767 	case 16: /* H265 */
768 		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
769 		image_size = ALIGN(image_size, 256);
770 
771 		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
772 		min_dpb_size = image_size * num_dpb_buffer;
773 		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
774 					   * 16 * num_dpb_buffer + 52 * 1024;
775 		break;
776 
777 	default:
778 		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
779 		return -EINVAL;
780 	}
781 
782 	if (width > pitch) {
783 		DRM_ERROR("Invalid UVD decoding target pitch!\n");
784 		return -EINVAL;
785 	}
786 
787 	if (dpb_size < min_dpb_size) {
788 		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
789 			  dpb_size, min_dpb_size);
790 		return -EINVAL;
791 	}
792 
793 	buf_sizes[0x1] = dpb_size;
794 	buf_sizes[0x2] = image_size;
795 	buf_sizes[0x4] = min_ctx_size;
796 	/* store image width to adjust nb memory pstate */
797 	adev->uvd.decode_image_width = width;
798 	return 0;
799 }
800 
801 /**
802  * amdgpu_uvd_cs_msg - handle UVD message
803  *
804  * @ctx: UVD parser context
805  * @bo: buffer object containing the message
806  * @offset: offset into the buffer object
807  *
808  * Peek into the UVD message and extract the session id.
809  * Make sure that we don't open up to many sessions.
810  */
811 static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
812 			     struct amdgpu_bo *bo, unsigned offset)
813 {
814 	struct amdgpu_device *adev = ctx->parser->adev;
815 	int32_t *msg, msg_type, handle;
816 	void *ptr;
817 	long r;
818 	int i;
819 
820 	if (offset & 0x3F) {
821 		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
822 		return -EINVAL;
823 	}
824 
825 	r = amdgpu_bo_kmap(bo, &ptr);
826 	if (r) {
827 		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
828 		return r;
829 	}
830 
831 	msg = ptr + offset;
832 
833 	msg_type = msg[1];
834 	handle = msg[2];
835 
836 	if (handle == 0) {
837 		DRM_ERROR("Invalid UVD handle!\n");
838 		return -EINVAL;
839 	}
840 
841 	switch (msg_type) {
842 	case 0:
843 		/* it's a create msg, calc image size (width * height) */
844 		amdgpu_bo_kunmap(bo);
845 
846 		/* try to alloc a new handle */
847 		for (i = 0; i < adev->uvd.max_handles; ++i) {
848 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
849 				DRM_ERROR(")Handle 0x%x already in use!\n",
850 					  handle);
851 				return -EINVAL;
852 			}
853 
854 			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
855 				adev->uvd.filp[i] = ctx->parser->filp;
856 				return 0;
857 			}
858 		}
859 
860 		DRM_ERROR("No more free UVD handles!\n");
861 		return -ENOSPC;
862 
863 	case 1:
864 		/* it's a decode msg, calc buffer sizes */
865 		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
866 		amdgpu_bo_kunmap(bo);
867 		if (r)
868 			return r;
869 
870 		/* validate the handle */
871 		for (i = 0; i < adev->uvd.max_handles; ++i) {
872 			if (atomic_read(&adev->uvd.handles[i]) == handle) {
873 				if (adev->uvd.filp[i] != ctx->parser->filp) {
874 					DRM_ERROR("UVD handle collision detected!\n");
875 					return -EINVAL;
876 				}
877 				return 0;
878 			}
879 		}
880 
881 		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
882 		return -ENOENT;
883 
884 	case 2:
885 		/* it's a destroy msg, free the handle */
886 		for (i = 0; i < adev->uvd.max_handles; ++i)
887 			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
888 		amdgpu_bo_kunmap(bo);
889 		return 0;
890 
891 	default:
892 		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
893 	}
894 
895 	return -EINVAL;
896 }
897 
898 /**
899  * amdgpu_uvd_cs_pass2 - second parsing round
900  *
901  * @ctx: UVD parser context
902  *
903  * Patch buffer addresses, make sure buffer sizes are correct.
904  */
905 static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
906 {
907 	struct amdgpu_bo_va_mapping *mapping;
908 	struct amdgpu_bo *bo;
909 	uint32_t cmd;
910 	uint64_t start, end;
911 	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
912 	int r;
913 
914 	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
915 	if (r) {
916 		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
917 		return r;
918 	}
919 
920 	start = amdgpu_bo_gpu_offset(bo);
921 
922 	end = (mapping->last + 1 - mapping->start);
923 	end = end * AMDGPU_GPU_PAGE_SIZE + start;
924 
925 	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
926 	start += addr;
927 
928 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
929 			    lower_32_bits(start));
930 	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
931 			    upper_32_bits(start));
932 
933 	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
934 	if (cmd < 0x4) {
935 		if ((end - start) < ctx->buf_sizes[cmd]) {
936 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
937 				  (unsigned)(end - start),
938 				  ctx->buf_sizes[cmd]);
939 			return -EINVAL;
940 		}
941 
942 	} else if (cmd == 0x206) {
943 		if ((end - start) < ctx->buf_sizes[4]) {
944 			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
945 					  (unsigned)(end - start),
946 					  ctx->buf_sizes[4]);
947 			return -EINVAL;
948 		}
949 	} else if ((cmd != 0x100) && (cmd != 0x204)) {
950 		DRM_ERROR("invalid UVD command %X!\n", cmd);
951 		return -EINVAL;
952 	}
953 
954 	if (!ctx->parser->adev->uvd.address_64_bit) {
955 		if ((start >> 28) != ((end - 1) >> 28)) {
956 			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
957 				  start, end);
958 			return -EINVAL;
959 		}
960 
961 		if ((cmd == 0 || cmd == 0x3) &&
962 		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
963 			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
964 				  start, end);
965 			return -EINVAL;
966 		}
967 	}
968 
969 	if (cmd == 0) {
970 		ctx->has_msg_cmd = true;
971 		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
972 		if (r)
973 			return r;
974 	} else if (!ctx->has_msg_cmd) {
975 		DRM_ERROR("Message needed before other commands are send!\n");
976 		return -EINVAL;
977 	}
978 
979 	return 0;
980 }
981 
982 /**
983  * amdgpu_uvd_cs_reg - parse register writes
984  *
985  * @ctx: UVD parser context
986  * @cb: callback function
987  *
988  * Parse the register writes, call cb on each complete command.
989  */
990 static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
991 			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
992 {
993 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
994 	int i, r;
995 
996 	ctx->idx++;
997 	for (i = 0; i <= ctx->count; ++i) {
998 		unsigned reg = ctx->reg + i;
999 
1000 		if (ctx->idx >= ib->length_dw) {
1001 			DRM_ERROR("Register command after end of CS!\n");
1002 			return -EINVAL;
1003 		}
1004 
1005 		switch (reg) {
1006 		case mmUVD_GPCOM_VCPU_DATA0:
1007 			ctx->data0 = ctx->idx;
1008 			break;
1009 		case mmUVD_GPCOM_VCPU_DATA1:
1010 			ctx->data1 = ctx->idx;
1011 			break;
1012 		case mmUVD_GPCOM_VCPU_CMD:
1013 			r = cb(ctx);
1014 			if (r)
1015 				return r;
1016 			break;
1017 		case mmUVD_ENGINE_CNTL:
1018 		case mmUVD_NO_OP:
1019 			break;
1020 		default:
1021 			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1022 			return -EINVAL;
1023 		}
1024 		ctx->idx++;
1025 	}
1026 	return 0;
1027 }
1028 
1029 /**
1030  * amdgpu_uvd_cs_packets - parse UVD packets
1031  *
1032  * @ctx: UVD parser context
1033  * @cb: callback function
1034  *
1035  * Parse the command stream packets.
1036  */
1037 static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1038 				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1039 {
1040 	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
1041 	int r;
1042 
1043 	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
1044 		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
1045 		unsigned type = CP_PACKET_GET_TYPE(cmd);
1046 		switch (type) {
1047 		case PACKET_TYPE0:
1048 			ctx->reg = CP_PACKET0_GET_REG(cmd);
1049 			ctx->count = CP_PACKET_GET_COUNT(cmd);
1050 			r = amdgpu_uvd_cs_reg(ctx, cb);
1051 			if (r)
1052 				return r;
1053 			break;
1054 		case PACKET_TYPE2:
1055 			++ctx->idx;
1056 			break;
1057 		default:
1058 			DRM_ERROR("Unknown packet type %d !\n", type);
1059 			return -EINVAL;
1060 		}
1061 	}
1062 	return 0;
1063 }
1064 
1065 /**
1066  * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1067  *
1068  * @parser: Command submission parser context
1069  * @ib_idx: Which indirect buffer to use
1070  *
1071  * Parse the command stream, patch in addresses as necessary.
1072  */
1073 int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
1074 {
1075 	struct amdgpu_uvd_cs_ctx ctx = {};
1076 	unsigned buf_sizes[] = {
1077 		[0x00000000]	=	2048,
1078 		[0x00000001]	=	0xFFFFFFFF,
1079 		[0x00000002]	=	0xFFFFFFFF,
1080 		[0x00000003]	=	2048,
1081 		[0x00000004]	=	0xFFFFFFFF,
1082 	};
1083 	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
1084 	int r;
1085 
1086 	parser->job->vm = NULL;
1087 	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
1088 
1089 	if (ib->length_dw % 16) {
1090 		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1091 			  ib->length_dw);
1092 		return -EINVAL;
1093 	}
1094 
1095 	ctx.parser = parser;
1096 	ctx.buf_sizes = buf_sizes;
1097 	ctx.ib_idx = ib_idx;
1098 
1099 	/* first round only required on chips without UVD 64 bit address support */
1100 	if (!parser->adev->uvd.address_64_bit) {
1101 		/* first round, make sure the buffers are actually in the UVD segment */
1102 		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1103 		if (r)
1104 			return r;
1105 	}
1106 
1107 	/* second round, patch buffer addresses into the command stream */
1108 	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1109 	if (r)
1110 		return r;
1111 
1112 	if (!ctx.has_msg_cmd) {
1113 		DRM_ERROR("UVD-IBs need a msg command!\n");
1114 		return -EINVAL;
1115 	}
1116 
1117 	return 0;
1118 }
1119 
1120 static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1121 			       bool direct, struct dma_fence **fence)
1122 {
1123 	struct amdgpu_device *adev = ring->adev;
1124 	struct dma_fence *f = NULL;
1125 	struct amdgpu_job *job;
1126 	struct amdgpu_ib *ib;
1127 	uint32_t data[4];
1128 	uint64_t addr;
1129 	long r;
1130 	int i;
1131 	unsigned offset_idx = 0;
1132 	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1133 
1134 	r = amdgpu_job_alloc_with_ib(adev, 64, direct ? AMDGPU_IB_POOL_DIRECT :
1135 				     AMDGPU_IB_POOL_DELAYED, &job);
1136 	if (r)
1137 		return r;
1138 
1139 	if (adev->asic_type >= CHIP_VEGA10) {
1140 		offset_idx = 1 + ring->me;
1141 		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1142 		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1143 	}
1144 
1145 	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1146 	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1147 	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1148 	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1149 
1150 	ib = &job->ibs[0];
1151 	addr = amdgpu_bo_gpu_offset(bo);
1152 	ib->ptr[0] = data[0];
1153 	ib->ptr[1] = addr;
1154 	ib->ptr[2] = data[1];
1155 	ib->ptr[3] = addr >> 32;
1156 	ib->ptr[4] = data[2];
1157 	ib->ptr[5] = 0;
1158 	for (i = 6; i < 16; i += 2) {
1159 		ib->ptr[i] = data[3];
1160 		ib->ptr[i+1] = 0;
1161 	}
1162 	ib->length_dw = 16;
1163 
1164 	if (direct) {
1165 		r = dma_resv_wait_timeout(bo->tbo.base.resv, true, false,
1166 					  msecs_to_jiffies(10));
1167 		if (r == 0)
1168 			r = -ETIMEDOUT;
1169 		if (r < 0)
1170 			goto err_free;
1171 
1172 		r = amdgpu_job_submit_direct(job, ring, &f);
1173 		if (r)
1174 			goto err_free;
1175 	} else {
1176 		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1177 				     AMDGPU_SYNC_ALWAYS,
1178 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1179 		if (r)
1180 			goto err_free;
1181 
1182 		r = amdgpu_job_submit(job, &adev->uvd.entity,
1183 				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1184 		if (r)
1185 			goto err_free;
1186 	}
1187 
1188 	amdgpu_bo_reserve(bo, true);
1189 	amdgpu_bo_fence(bo, f, false);
1190 	amdgpu_bo_unreserve(bo);
1191 
1192 	if (fence)
1193 		*fence = dma_fence_get(f);
1194 	dma_fence_put(f);
1195 
1196 	return 0;
1197 
1198 err_free:
1199 	amdgpu_job_free(job);
1200 	return r;
1201 }
1202 
1203 /* multiple fence commands without any stream commands in between can
1204    crash the vcpu so just try to emmit a dummy create/destroy msg to
1205    avoid this */
1206 int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1207 			      struct dma_fence **fence)
1208 {
1209 	struct amdgpu_device *adev = ring->adev;
1210 	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1211 	uint32_t *msg;
1212 	int i;
1213 
1214 	msg = amdgpu_bo_kptr(bo);
1215 	/* stitch together an UVD create msg */
1216 	msg[0] = cpu_to_le32(0x00000de4);
1217 	msg[1] = cpu_to_le32(0x00000000);
1218 	msg[2] = cpu_to_le32(handle);
1219 	msg[3] = cpu_to_le32(0x00000000);
1220 	msg[4] = cpu_to_le32(0x00000000);
1221 	msg[5] = cpu_to_le32(0x00000000);
1222 	msg[6] = cpu_to_le32(0x00000000);
1223 	msg[7] = cpu_to_le32(0x00000780);
1224 	msg[8] = cpu_to_le32(0x00000440);
1225 	msg[9] = cpu_to_le32(0x00000000);
1226 	msg[10] = cpu_to_le32(0x01b37000);
1227 	for (i = 11; i < 1024; ++i)
1228 		msg[i] = cpu_to_le32(0x0);
1229 
1230 	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1231 
1232 }
1233 
1234 int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1235 			       bool direct, struct dma_fence **fence)
1236 {
1237 	struct amdgpu_device *adev = ring->adev;
1238 	struct amdgpu_bo *bo = NULL;
1239 	uint32_t *msg;
1240 	int r, i;
1241 
1242 	if (direct) {
1243 		bo = adev->uvd.ib_bo;
1244 	} else {
1245 		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1246 		if (r)
1247 			return r;
1248 	}
1249 
1250 	msg = amdgpu_bo_kptr(bo);
1251 	/* stitch together an UVD destroy msg */
1252 	msg[0] = cpu_to_le32(0x00000de4);
1253 	msg[1] = cpu_to_le32(0x00000002);
1254 	msg[2] = cpu_to_le32(handle);
1255 	msg[3] = cpu_to_le32(0x00000000);
1256 	for (i = 4; i < 1024; ++i)
1257 		msg[i] = cpu_to_le32(0x0);
1258 
1259 	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1260 
1261 	if (!direct)
1262 		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1263 
1264 	return r;
1265 }
1266 
1267 static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1268 {
1269 	struct amdgpu_device *adev =
1270 		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1271 	unsigned fences = 0, i, j;
1272 
1273 	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1274 		if (adev->uvd.harvest_config & (1 << i))
1275 			continue;
1276 		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1277 		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1278 			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1279 		}
1280 	}
1281 
1282 	if (fences == 0) {
1283 		if (adev->pm.dpm_enabled) {
1284 			amdgpu_dpm_enable_uvd(adev, false);
1285 		} else {
1286 			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1287 			/* shutdown the UVD block */
1288 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1289 							       AMD_PG_STATE_GATE);
1290 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1291 							       AMD_CG_STATE_GATE);
1292 		}
1293 	} else {
1294 		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1295 	}
1296 }
1297 
1298 void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1299 {
1300 	struct amdgpu_device *adev = ring->adev;
1301 	bool set_clocks;
1302 
1303 	if (amdgpu_sriov_vf(adev))
1304 		return;
1305 
1306 	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1307 	if (set_clocks) {
1308 		if (adev->pm.dpm_enabled) {
1309 			amdgpu_dpm_enable_uvd(adev, true);
1310 		} else {
1311 			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1312 			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1313 							       AMD_CG_STATE_UNGATE);
1314 			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1315 							       AMD_PG_STATE_UNGATE);
1316 		}
1317 	}
1318 }
1319 
1320 void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1321 {
1322 	if (!amdgpu_sriov_vf(ring->adev))
1323 		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1324 }
1325 
1326 /**
1327  * amdgpu_uvd_ring_test_ib - test ib execution
1328  *
1329  * @ring: amdgpu_ring pointer
1330  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1331  *
1332  * Test if we can successfully execute an IB
1333  */
1334 int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1335 {
1336 	struct dma_fence *fence;
1337 	long r;
1338 
1339 	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1340 	if (r)
1341 		goto error;
1342 
1343 	r = dma_fence_wait_timeout(fence, false, timeout);
1344 	dma_fence_put(fence);
1345 	if (r == 0)
1346 		r = -ETIMEDOUT;
1347 	if (r < 0)
1348 		goto error;
1349 
1350 	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1351 	if (r)
1352 		goto error;
1353 
1354 	r = dma_fence_wait_timeout(fence, false, timeout);
1355 	if (r == 0)
1356 		r = -ETIMEDOUT;
1357 	else if (r > 0)
1358 		r = 0;
1359 
1360 	dma_fence_put(fence);
1361 
1362 error:
1363 	return r;
1364 }
1365 
1366 /**
1367  * amdgpu_uvd_used_handles - returns used UVD handles
1368  *
1369  * @adev: amdgpu_device pointer
1370  *
1371  * Returns the number of UVD handles in use
1372  */
1373 uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1374 {
1375 	unsigned i;
1376 	uint32_t used_handles = 0;
1377 
1378 	for (i = 0; i < adev->uvd.max_handles; ++i) {
1379 		/*
1380 		 * Handles can be freed in any order, and not
1381 		 * necessarily linear. So we need to count
1382 		 * all non-zero handles.
1383 		 */
1384 		if (atomic_read(&adev->uvd.handles[i]))
1385 			used_handles++;
1386 	}
1387 
1388 	return used_handles;
1389 }
1390